A swarm-inspired re-ranker system for statistical machine translation

Candidate re-ranking for SMT-based grammatical error correction

“…Various global features were investigated for SMT re-ranking, such as the decoder's scores, source and target sentences, alignments and POS tags, sentence type probabilities, posterior probabilities and back translation features. More recently, Farzi and Faili (2015) proposed a re-ranking system based on swarm algorithms.…”

Improving Arabic neural machine translation via n-best list re-ranking

“…Experiments with English-to-Portuguese translation showed a significant improvement that varied between 1.5 and 2.5 absolute BLEU points. Farzi and Faili (2015) used a set of non-syntactical features to re-rank the n-best translation candidates generated by a Phrase-based Statistical Machine Translation system. They investigated several feature weights optimization algorithms such as Particle Swarm Optimization (PSO), Quantum-behaved Particle Swarm Optimization (QPSO), Genetic Algorithms (GA), Perceptron and Averaged Perceptron.…”

“…The features set that we have proposed covers the lexical, syntactic and semantic aspects of the translation candidates (the n-best list candidates) and can be grouped into five classes: (1) Translation-based Features: these features are related to a translation model and can be helpful for promoting translation adequacy; (2) Fluency Features: these features are used to promote syntactic fluency by incorporating language models; (3) Length-based Features: these features are used to promote the n-best list candidates according to the likelihood of their lengths; (4) N-best list Features: these features are extracted directly from the n-best list and are used to promote the most likely candidates in it; and (5) Embedding Features: these are based on bilingual word embeddings and are used to cover the semantic aspect of the translation candidates. For the problem of feature weights optimization, we present a methodology that is fairly similar to the one presented by Farzi and Faili (2015). It uses a Quantum-behaved Particle Swarm Optimization (QPSO) which guarantees the global convergence of the optimization process.…”

“…It uses a Quantum-behaved Particle Swarm Optimization (QPSO) which guarantees the global convergence of the optimization process. Though our overall re-ranking framework is closely related to the one proposed by Farzi and Faili (2015), we diverge from them in various aspects:…”

“…-A new class of embedding-based features is proposed to take into account the semantic aspects of the translation candidates via bilingual word embeddings. -Unlike Farzi and Faili (2015), our proposed objective function is based on the corpuslevel, not the sentence-level, BLEU Score. Thus, feature weights are updated only if an improvement is achieved on the whole development corpus (section 4.2).…”